DE3587251T2 - ADAPTABLE METHOD AND DEVICE FOR VOICE CODING. - Google Patents

Description

The invention relates to a speech coder as specified in the preamble of claim 1. A speech coder of this type is known from WP-8-0 176 243.

In the above-mentioned document, an encoder for speech signals is disclosed which comprises separating means for receiving speech signals and generating series of values, each series representing respective portions of the frequency spectrum of the input signal, and encoding means for digitally encoding each series, and bit detection means for changing the number of bits used to encode the respective series depending on their relative energy content, the number of series to which any given number of bits is assigned being constant and only the selection of the series to which respective numbers of bits are assigned being changed.

Conventional analog telephone systems are being replaced by digital systems. In digital systems, the analog signals are sampled at a rate of about twice the bandwidth of the analog signals, or about 8 kilohertz, and the samples are then encoded. In a simple pulse code modulation (PCN) system, each sample is quantized as one of a certain set of preselected values and encoded as a digital word, which is then transmitted over the telephone lines. For example, with digital words of 8 bits, the analog sample is quantized to 28, or 256 levels, each of which is denoted by a different word of 8 bits. When nonlinear quantization is used, excellent quality speech can be transmitted using only 7 bits per sample. can be obtained; however, since a 7-bit word is still required for each sample, bit transfer rates of 56 kilobits per second are necessary.

Efforts have been made to reduce the bit rates required to encode speech and obtain a clear, decoded speech signal at the receiving end of the system. The linear predictive coding technique (LPC) is based on the recognition that speech production involves an excitation and a filtering process. The excitation is determined by the vocal cord vibration for spoken speech and by turbulence for non-speech speech and this excitation signal is then modified by the filtering process of the vocal resonant chambers including the mouth and nose passages. For a particular group of samples, a digital filter simulating the formant effects of the resonant chambers can be defined and the definition can be coded. A residual signal approximating the excitation can then be obtained by passing the speech signal through an inverted formant filter and the residual signal can be coded. Since sufficient information is contained in the low frequency portion of the residual spectrum, it is possible to encode only the low frequency baseband and still obtain sufficiently clear speech. At the receiver, a definition of the formant filter and the residual baseband are decoded. The baseband is repeated to complete the spectrum of the residual signal. By applying the decoded filter to the repeated baseband signal, the source speech can be reconstructed.

A significant problem with the predictive coding (LPC) approach is defining the formant filter, which must be redefined with each window from above. A complex encoder and a complex decoder are required to obtain transmission rates as low as 16,000 bits per second. Another problem with such systems is that they do not always provide a satisfactory reconstruction of certain formants, such as that resulting from nasal resonance. It is the object of the invention to solve these difficulties.

This objective is achieved by the characterizing features of claim 1 and 14. Preferred embodiments of the invention are the subject of the subclaims.

In one system, the approximate envelope of the transform spectrum is defined in each of a plurality of subbands of the coefficients, and each definition of the envelope is encoded for transmission. Each spectral coefficient is then scaled relative to the defined envelope of the corresponding subband, and each scaled coefficient is encoded in a number of bits determined by the defined envelope of its subband.

Zero bits may be allocated to a number of less significant subbands as indicated by the defined envelopes; and variable numbers of bits may be used for each coded coefficient depending on the size of the defined envelope for the respective subband. Thus, the subbands that are transmitted and the resolution with which the transmitted subbands are encoded are adaptively determined for each sampling window based on the defined envelopes of the subbands.

At the receiver, the subbands that have been transmitted are repeated to define coefficients of the frequencies that have not been transmitted. A list repetition procedure is followed by which an nth coefficient that has been transmitted is repeated as an nth coefficient that has not been transmitted. After repetition, the speech signal can be regenerated by using the transmitted envelope definitions used to inversely scale coefficients of the respective subbands and by performing an inverse transformation.

In another system, the spectrum is first normalized with respect to only a few bins and subsequently with respect to a larger number of subbins. The maximum magnitude in each of the bins and in each of the subbins is encoded. The maxima are logarithmically encoded and only a baseband of the normalized spectrum is encoded.

The foregoing and other objects, features and advantages of the invention will become apparent from the following more particular description of a preferred embodiment of the invention as illustrated in the accompanying drawings, in which like reference characters refer to like parts throughout the several views. The drawings are not necessarily to scale, emphasis instead being placed upon illustrating the principles of the invention.

Fig. 1 is a block diagram representation of an encoder and a decoder embodying the invention;

Fig. 2 is a block diagram of a speech encoder and a corresponding decoder of a preferred embodiment of the system of Fig. 1.

Fig. 3 is an example of a magnitude spectrum of the Fourier transform of a speech window illustrating the principles of the system of Fig. 2;

Fig. 4 is an example spectrum normalized from Fig. 3 based on the principles of present invention;

Fig. 5 shows schematically a quantization device for complex values of the normalized spectrum;

Figure 6 is an example representation of coefficient groups being transmitted and shows the repetition technique of the system of Figure 2.

Fig. 7 is an example of a range of sizes of a speech window illustrating the principles of another system embodying the present invention;

Fig. 8 is an example spectrum normalized from the spectrum of Fig. 7 using four formant sections;

Fig. 9 is an example spectrum normalized from that of Fig. 8 into subbands;

Fig. 10 schematically illustrates a quantization device for complex values of the normalized spectrum ;

Figure 11 is a block diagram representation of the spectral equalization coding circuit of Figure 1 in an alternative embodiment.

A block diagram of the system is shown in Fig. 1. The speech is filtered with a telephone bandpass filter 20, which prevents pseudonymization as the signal is sampled 8,000 times per second in a sampling circuit 22. The analog samples are digitally encoded in an analog-to-digital encoder 24 and preprocessed at 26 before being passed to a discrete Fourier transform unit 28.

The output of the Fourier transform circuit 28 is a series of coefficients indicating the magnitude and phase of the Fourier transform spectrum at each of 97 frequencies spaced 41,667 hertz apart. The magnitude spectrum of the Fourier transform output is shown in Fig. 3 as a continuous function, but it can be seen that the transform circuit 28 actually provides only 97 stepwise outputs.

According to the present invention, the Fourier transform spectrum of the full speech is equalized within a selected window and encoded in a circuit 30 in a manner to be discussed below. The resulting digital signal can be transmitted to a receiver at 16,000 bits per second over a line 32. At the receiver, the full spectrum of Figure 3 is reconstructed in the circuit 34. The inverse Fourier transform is performed in the circuit 36 and applied to a post-processor 38 corresponding to the pre-processor 26. This signal is then converted to analog form in a digital-to-analog converter 40. Final filtering in the filter 42 provides clear speech to the listener.

A preferred system uses a pipelined multiprocessor architecture. One microcomputer is dedicated to analog-to-digital conversion with pre-emphasis filters, one is dedicated to forward Fourier transform, and a third is dedicated to spectral equalization and coding. Similarly, in the receiver, one microcomputer is dedicated to spectral reconstruction, another to inverse Fourier transform, and a third to digital-to-analog conversion with post-emphasis filters.

The spectral equalization and coding technique of the present invention is based on the recognition that the Fourier transform of the total signal has a relatively flat spectrum of pitch shown in Fig. 4 and shaped by the formant signals. In the present system, the signal of Fig. 4 is obtained by normalizing the spectrum of Fig. 3 to at least one curve which can itself be coded separately from the residual spectrum of Fig. 4.

An embodiment of the coding system of Fig. 1 is shown in Fig. 2. Before compression, the analog speech signal is low-pass filtered in filter 20 at 3.4 kilohertz, sampled in sampler 22 at a rate of 8 kilohertz, and digitized using a 12-bit linear analog-to-digital converter 24. It will be appreciated that the input to the encoder may already be in digital form and requires conversion to the code that can be accepted by the encoder. The digital speech signal in frames of N samples is first scaled up in a scaling device 26 to maximize its dynamic range in each frame. The scaled input samples are then Fourier transformed in a fast Fourier transform facility 28 to obtain a corresponding discrete spectrum represented by (N/2)+1 complex frequency coefficients.

In a particular embodiment, the input frame size is equal to 180 samples, corresponding to one frame every 22.5 milliseconds. However, the discrete Fourier transform is performed on 192 samples, including 12 samples overlapped with the previous frame, preceded by trapezoidal windowing with a slope of 12 points at each end. The resulting fast Fourier transform output includes 97 complex frequency coefficients spaced 41,667 hertz apart.

An example of a Fourier transform magnitude spectrum output by the fast Fourier transform 28 is shown in Fig. 3. Although it is shown as a continuous function, it can be seen that the transformation circuit 28 actually provides only 97 stepwise, complex outputs.

The magnitude spectrum of the Fourier transform output is equalized and encoded. To do this, the spectrum is divided into adjacent subbands and an estimated spectral envelope is based on a piecewise approximation of these subbands at 44. In a particular implementation, the spectrum is divided into twenty subbands, each of which includes four complex coefficients. Frequencies above 3,291.67 Hz are not encoded and are set equal to zero at the receiver. To equalize the spectrum, the spectral envelope of each subband is assumed to be constant and is defined by the peak magnitude in each subband, as shown by the horizontal lines in Figure 3. Each magnitude, or more precisely its inverse, can be treated as a scaling factor for its respective subband. Each scaling factor is quantized in a quantization device 45 with four bits.

Then, by multiplying the magnitude of each coefficient of the spectrum by the scaling factor associated with that coefficient at 46, the residual flattened spectrum of Fig. 4 is obtained. This flattening of the spectrum is equivalent to reverse filtering of the signal based on the piecewise constant estimate of the spectral envelope.

Only selected subbands of the flattened spectrum of Fig. 4 are quantized and transmitted. The selection of 48 subbands to be transmitted is based on the scaling factor of the subbands. In a particular embodiment, the twelve subbands that have the smallest scaling factors, i.e. the highest energy, are encoded and For the 8 low energy bands, only the scaling factors are transmitted.

A non-uniform bit allocation is used for the complex coefficients that are transmitted. Three separate two-dimensional quantizers 50 are used for the 12 subbands transmitted. The sixteen complex coefficients for the four subbands that have the smallest scale factors are each quantized with seven bits. The coefficients of the four subbands that have the next smallest scale factors are each quantized with 6 bits, and the coefficients of the remaining four of the transmitted subgroups are each quantized with 4 bits. In fact, the coefficients of the eight subbands that are not transmitted are quantized with zero bits.

Each of the two-dimensional quantizers is designed using an approach given by Linde et al., "An Algorithm for Vector Quantizer Design", IEEE Trans on Commun, Vol COM-28, pp. 84-95, January 1980. The result for the 7-bit quantizer is shown in Fig. 5. The two dimensions of the quantizer are the real and imaginary parts of each complex coefficient. Each bundle has a 7-bit representation to which each complex point in the bundle is quantized. The actual quantization can be done by a look-up table in read-only memory.

The bit allocation for a single frame can be summarized as follows:

Scaling factors 20 · 4 bits each = 80 bits

16 · 7 bits = 112 bits

16 · 6 bits = 96 bits

16 · 4 bits = 64 bits

Time scaling = 4 bit

Synchronization = 4 bits

Total = 360 bits

At the receiver, the transmitted twelve groups of coefficients are applied to respective inverse quantizers of seven bits, six bits and four bits at 52. The frequency subbands to which the resulting coefficients correspond are determined by the scale factors transmitted in turn for all subbands. Thus, the coefficients from the seven bit inverse quantizer are arranged in the subbands of which the scale factors indicate which has the largest magnitude.

The coefficients of the eight subbands that have not been transmitted are approximated by repeating the transmitted subbands at 54. A list repetition approximation is used for this. This approximation is shown in Figure 6. In Figure 6, the coefficients for each subband are given by a single vector. The transmitted subbands are denoted T1, T2, T3,... Tn,... and the subbands that must be generated by repetition in the receiver are denoted R1, R2, R3,... Rn,... According to the repetition technique of the present system, the coefficients of subband Tn are used for both Tn and Rn. Thus, the scaled coefficients for subband T1 are repeated at subband R1, those of subband T2 are repeated at R2, and those of subband T3 are repeated at R3. The reason for this list repetition technique is that the subbands themselves are usually grouped into blocks of transmitted subbands and blocks of non-transmitted subbands. Thus, large blocks of coefficients are typically repeated using this method, and the speech harmonics are maintained in the repetition process.

Once the equalized spectrum of Figure 4 has been recreated by repeating the subbands, a reproduction of the spectrum of Figure 3 can be created at 56 by applying the scaling factors to the equalized spectrum. From this reproduction of the Fourier transform of the original Fourier transform, the speech can be obtained by an inverse fast Fourier transformer 36, an inverse counter 38, a digital-to-analog converter 40 and a reproduction filter 42.

An outstanding advantage of the present invention is that the encoder is not based on an assumed fixed low-pass spectral model that is speech specific. Voiceband data and voiceband signaling take the form of sine waves of some bandwidth that may occur at any frequency. Where only a lower and an upper baseband of coefficients are transmitted, voiceband data may be lost. In the present system, the subbands in which the digital information is transmitted are naturally selected for their higher energy.

Another attractive feature of the coding system is its integrated data rate coding capability. Integrated coding, which is important as a method of congestion control in telephone applications, allows the data to leave the encoder at a constant bit rate but be received at the decoder at a lower bit rate because some bits are removed along the way. Integrated coding involves a packet or block of bits in which there is a hierarchy of sub-blocks. The least critical sub-blocks may be removed first when the channel becomes congested. This hierarchical principle is a natural one in the present system, in which the sub-band information described by a set of frequency coefficients is ordered in decreasing order of importance. and the missing coefficients can always be approximated by the received ones. The more coefficients there are in the sentence, the higher the rate and the better the quality. However, the speech quality decreases very smoothly with a small decrease in the rate. The incorporation of an integrated coding system in conjunction with this method is therefore quite simple and very attractive.

The coding technique described above provides excellent speech coding and reproduction at 16 kilobits per second. Excellent results, as low as 8.0 kilobits per second, can be obtained by using this technique in conjunction with a frequency scaling technique known as time domain harmonic scaling, described by D. Malah, "Time Domain Algorithm for Harmonic Bandwidth Reduction and Time Scaling of Speech, Signals," IEEE Trans. Acoust., Speech, Signal Processing, Vol. ASSP-27, pp. 121-133, April 1979. In this technique, speech at twice the rate of the original speech but at the original pitch is generated by combining adjacent pitch cycles before performing the fast Fourier transform. The frequency-scaled speech can then be more quickly Fourier transformed in the technique described above.

Although each of the residue extraction, subband selection and quantization steps and the reverse quantization, repetition and envelope excitation steps are shown as individual elements of the system, it will be appreciated that they can be integrated into an actual system. For example, the residue spectrum need not be obtained for subbands that are not transmitted. The system can then be implemented using a combination of software and hardware.

In another coding system, the shape of the spectrum by a two-step process. This process also encodes the shape of the entire spectrum from 100 to 3800 Hz, as this is convenient for baseband coding. In the first step, the spectrum is divided into four sections, which are shown in Fig. 7.

125 - 583 Hz

625 - 1959 Hz

2000 - 3416 Hz

3468 - 3833 Hz

These bins roughly correspond to the location of the first four formants. The dynamic range of the spectral coefficient magnitudes is much smaller within each of these bins than in the spectrum as a whole. For pitch phenomena, the peak magnitude near 250 Hz may be 30 dB above the magnitudes near 3800 Hz. The first step of spectral normalization is performed by finding the peak magnitudes within each bin, quantizing these peaks to 5 bits each with a logarithmic quantizer, and dividing each spectral coefficient by the quantized peak in that bin. The result is a vector of spectral coefficients with maximum magnitude equal to unity. The division into bins should result in the spectral coefficients being reasonably uniformly distributed within the complex disk of radius one.

The second step gains a more detailed structure. The spectrum is divided into equal bands of approximately 165 Hz each. The peak magnitude within each band is detected and quantized to 3 bits. The complex spectral coefficients within the band are divided by the quantized magnitude and encoded to 6 bits each using a hexagonal quantizer. This encoding preserves the phase information, which is important for recovering frame boundaries.

The details of this alternative approach are shown with reference to Figures 7 through 11. In this system, the preprocessor 26 is a single pole preemphasis filter. Low frequencies are attenuated by approximately 5 dB. High frequencies are boosted. The highest frequency (4 kHz) is emphasized by approximately 24 dB. The filter is useful in equalizing the spectrum by reducing the low pass effects of the initial filter and the high frequency attenuation of the lips. The emphasis helps maintain numerical accuracy in the subsequent calculation of the Fourier transform.

In each of the four formant sections, the spectrum is normalized to a curve, which in this case is chosen to be a horizontal line through the peak magnitude of the spectrum in each region. These curves are shown as lines 58, 60, 62 and 64 in Figure 7. The peak magnitude of the complex numbers in each section is determined and encoded to five bits at unit 66 of Figure 11 by finding a value k which is encoded such that the peak value is between 162 x 212(k-1)/32 and 162 x 212k/32. This gives a logarithmic encoding of the peak magnitude. The four k values, each encoded to 5 bits, give a total of 20 bits from the formant encoder, which are the most significant bits of the transmitted code for the window. All spectral coefficients in each of the four regions are then divided by 162 212k/32 in the spectral normalization unit 68. With this procedure, all resulting quantities shown in Fig. 8 are less than 1.

Next, the normalized coefficients output from unit 68 are grouped into twenty ranges of four and two subranges of five, shown in Fig. 8. The peak size in each of these subranges is determined and encoded to 3 bits with a logarithmic quantizer in unit 70. The peak is always encoded to the next largest value. The three bits from each of the 22 subranges provide an additional 66 bits of final signal for the window. Each output within a subrange is multiplied by the reciprocal of the quantized magnitude in the sample normalization unit 72, thereby ensuring that all outputs remain less than 1, as shown in Fig. 9.

Each complex output from the baseband of 125 Hz to 1959 Hz of the normalized spectrum of Figure 9 is encoded to 6 bits by the two-dimensional quantizer and encoder 74. The two-dimensional quantizer is formed by dividing a complex disk of radius one into hexagons as shown in Figure 10. The x, y coordinates are radially deformed by an exponential function to approximate a logarithmic encoding of magnitude. All points within a hexagon are quantized to the coordinates of the center of the hexagon. As a result, larger magnitude coefficients are encoded with better phase resolution than smaller magnitude coefficients. The actual quantization is performed by a look-up table, but efficient computational algorithms are possible.

The bit allocation for a single frame can be summarized as follows:

Formant range scaling factors 4 · 5 bits each = 20 bits

Subband scaling factors 22 · 3 bits each = 66 bits

Baseband components 45 · 6 bits each = 270 bits

Total 356 bit

For a practical 16 Kb/s transmission system, this allows 4 bits per frame for additional functions such as frame synchronization. The actual coding transforms, bit allocations and subband sizes can be changed as the encoder is optimized for different applications.

All normalization factors (four of 5 bits each, 22 of three bits each) and the encoded, normalized baseband coefficients (45 of 6 bits) are transmitted. At the receiver, the baseband is decoded and copied to the upper frequency range. The normalization factors are applied to the spectrum to restore the original shape. In particular, in the receiver, inputs 0 to 2 and 92 to 96 are set to zero for the inverse Fourier transform. The normalized complex coefficients for inputs 3 to 47 are reconstructed from the quantization codes by table lookup. They are copied to locations 48 to 92. This copying is the nonlinear recovery step. The scaling factors for the subranges and the larger ranges are then applied.

The inverse transform is calculated in unit 36. The effects of windowing are removed by adding the last twelve points of the previous inverse transform to the first twelve points of the current inverse transform. The speech now passes through filter 38, which is an inverse of the preemphasis filter and which attenuates the high frequencies, removes the effects of high pitch boost and reduces high frequency quantization noise. The outputs are converted to analog using a 12-bit linear analog-to-digital converter 40.

The baseband that is repeated in spectrum reconstruction has been described as a band of lower frequencies. However, the baseband can include any frequency range within the spectrum. For some sounds, where higher energy levels are found at the higher frequencies, a higher frequency baseband is preferred.

It should be noted that the baseband deteriorates suffers only from quantization errors. The recovery of the upper frequencies is only as good as the model and the shaping information. However, by ensuring that at least some coefficient in each 165 Hz band of the normalized baseband is at its natural magnitude, each formant is excited at approximately the correct frequency. This is an improvement over baseband residual excitation, where some parts of the spectrum may be under-energetic. The reduction in computational complexity due to finding and scaling the peak rather than linear prediction analysis and filtering is very significant.

This method is a wideband approach because the entire pitch frequency range is encoded. The main difficulty with other wideband systems at 16 kb/s is that there are hardly enough bits available to give a rough description of the waveform. Baseband excitation systems, such as the present system, deal with this difficulty by devoting most of the bits to the baseband and restoring the excitation signal for higher frequencies. A modification of the subband coding transformation just described could encode the baseband as described above, but encode only some energy values for the higher frequencies. Frequency translation of the baseband recreates the fine structure of the upper spectrum.

While the invention has been particularly shown and described with reference to a preferred embodiment, those of ordinary skill in the art will recognize that various changes in form and details may be made without departing from the scope of the invention as defined by the appended claims.

Claims (18)

1. Speech coder comprising: a Fourier transform device (28) for performing a discrete Fourier transform of an incoming speech signal to produce a discrete transform spectrum of coefficients; a normalization device (30) for modifying the transform spectrum to produce a normalized, flatter spectrum and for encoding a function by which the discrete spectrum is modified; and a device (30) for coding at least part of the spectrum, characterized in that the normalization means (30) comprises means (44) for defining the approximate envelope of the discrete spectrum in each of a plurality of subbands of coefficients and for encoding the defined envelope of each subband of coefficients and means for scaling each spectrum coefficient relative to the defined envelope of the respective subband of coefficients; and the means (30) for encoding which encodes the scaled spectrum coefficients within each subband into a number of bits determined by the defined envelope of the subband. 2. A speech coding system according to claim 1, wherein the number of bits determined for a plurality of subbands is zero, so that the scaled coefficients for those subbands are not transmitted. 3. A speech coding system according to claim 2, wherein the scaling coefficients of different subbands are encoded into different numbers of bits other than zero. 4. A speech coding system according to claim 2, wherein the coded speech is decoded by repeating subbands of transmitted coefficients as replacements for subbands of non-transmitted coefficients, the transmitted coefficients being repeated such that the transmitted nth subband is repeated as the non-transmitted nth subband. 5. A speech coding system according to claim 1, wherein the coefficients of different subbands are encoded into different numbers of bits which are not equal to zero. 6. A speech coding system according to claim 1, wherein the means (30) for coding encodes the scaled coefficients of less than all of the subbands, the encoded scaled coefficients being those corresponding to the defined envelopes of larger amplitude, the scaled coefficients of subbands corresponding to the defined envelopes of largest amplitudes being encoded in more bits than the coefficients of subbands corresponding to the defined envelopes of smaller amplitudes. 7. A speech coding system according to claim 6, wherein the coded speech is decoded by repeating subbands of transmitted coefficients as replacements for subbands of non-transmitted coefficients, the transmitted coefficients being repeated such that the transmitted nth subband is repeated as the non-transmitted nth subband. 8. Speech coding system according to claim 6, wherein the transforming means (28) carries out a discrete Fourier transform. 9. A speech coding system according to claim 1, wherein the normalization means includes: means (44) for determining the maximum amplitude of the discrete spectrum within each of a plurality of regions of the spectrum; and means for digitally encoding the maximum amplitude of each range; and means (45) for scaling each coefficient of the discrete spectrum in each region to the maximum amplitude of each region to produce a first set of normalized coefficients. 10. A speech coding system according to claim 9, wherein the normalization means further includes: means for determining the maximum amplitude of the first set of normalized outputs in each of a plurality of sub-regions of the spectrum; Means for digitally encoding the maximum amplitude of each subrange; and Means for scaling each output of the first set of normalized outputs to the maximum amplitude of each subrange to produce a second set of normalized outputs. 11. A speech coder according to claim 10, wherein each of the maximum amplitudes is logarithmically encoded. 12. A speech coder according to claim 10, wherein the maximum amplitude is determined for each of four regions corresponding to the first four formats. 13. Speech coder according to claim 10, wherein only a baseband of the normalized spectrum is coded. 14. A method for speech coding, comprising: Performing a discrete Fourier transform of a speech window to produce a discrete transform spectrum; generating a normalized spectrum by defining at least one curve that approximates the amplitude of the discrete spectrum, digitally encoding the defined curve and defining the discrete spectrum with respect to the defined curve; and Coding at least a portion of the standardized spectrum, characterized in that the normalized spectrum is generated by defining the approximate envelope of the discrete spectrum in each of a plurality of subbands of coefficients and digitally encoding the defined envelope of each subband of coefficients and scaling each coefficient with respect to the defined amplitude of the respective subband of coefficients; and the scaled coefficients within each subband are encoded into a number of bits determined by the defined envelope of the subband. 15. A method according to claim 14, wherein the number of bits determined for a plurality of subbands is zero, so that the scaled coefficients for those subbands are not transmitted. 16. The method of claim 15, wherein the scaled coefficients of different subbands are encoded into different numbers of non-zero bits. 17. A method according to claim 15, wherein the coded speech is decoded by repeating subbands of transmitted coefficients as replacements for subbands of non-transmitted coefficients, the transmitted coefficients being repeated such that the transmitted nth subband is repeated by the non-transmitted nth subband. 18. The method of claim 14, wherein the normalized spectrum is generated by: determining a maximum amplitude of the discrete spectrum within each of a plurality of regions of the spectrum; digital coding of the maximum amplitude of each range; and Scale each coefficient of the discrete spectrum in each range to the maximum amplitude of each range to produce a set of normalized coefficients.